Re: [PATCH v9 4/5] locking/qspinlock: Introduce starvation avoidance into CNA

[Waiman Long <longman@redhat.com>]

On 1/14/20 10:59 PM, Alex Kogan wrote:
> Keep track of the number of intra-node lock handoffs, and force
> inter-node handoff once this number reaches a preset threshold.
> The default value for the threshold can be overridden with
> the new kernel boot command-line option "numa_spinlock_threshold".
>
> Signed-off-by: Alex Kogan <alex.kogan@oracle.com>
> Reviewed-by: Steve Sistare <steven.sistare@oracle.com>
> Reviewed-by: Waiman Long <longman@redhat.com>
> ---
>  .../admin-guide/kernel-parameters.txt         |  8 ++++
>  kernel/locking/qspinlock.c                    |  3 ++
>  kernel/locking/qspinlock_cna.h                | 41 ++++++++++++++++++-
>  3 files changed, 51 insertions(+), 1 deletion(-)
>
> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
> index b68cb80e477f..30d79819a3b0 100644
> --- a/Documentation/admin-guide/kernel-parameters.txt
> +++ b/Documentation/admin-guide/kernel-parameters.txt
> @@ -3200,6 +3200,14 @@
>  			Not specifying this option is equivalent to
>  			numa_spinlock=auto.
>  
> +	numa_spinlock_threshold=	[NUMA, PV_OPS]
> +			Set the threshold for the number of intra-node
> +			lock hand-offs before the NUMA-aware spinlock
> +			is forced to be passed to a thread on another NUMA node.
> +			Valid values are in the [0..31] range. Smaller values
> +			result in a more fair, but less performant spinlock, and
> +			vice versa. The default value is 16.
> +
>  	cpu0_hotplug	[X86] Turn on CPU0 hotplug feature when
>  			CONFIG_BOOTPARAM_HOTPLUG_CPU0 is off.
>  			Some features depend on CPU0. Known dependencies are:
> diff --git a/kernel/locking/qspinlock.c b/kernel/locking/qspinlock.c
> index 609980a53841..e382d8946ccc 100644
> --- a/kernel/locking/qspinlock.c
> +++ b/kernel/locking/qspinlock.c
> @@ -597,6 +597,9 @@ EXPORT_SYMBOL(queued_spin_lock_slowpath);
>  #if !defined(_GEN_CNA_LOCK_SLOWPATH) && defined(CONFIG_NUMA_AWARE_SPINLOCKS)
>  #define _GEN_CNA_LOCK_SLOWPATH
>  
> +#undef pv_init_node
> +#define pv_init_node			cna_init_node
> +
>  #undef pv_wait_head_or_lock
>  #define pv_wait_head_or_lock		cna_pre_scan
>  
> diff --git a/kernel/locking/qspinlock_cna.h b/kernel/locking/qspinlock_cna.h
> index 8000231f3d51..a2b65f87e6f8 100644
> --- a/kernel/locking/qspinlock_cna.h
> +++ b/kernel/locking/qspinlock_cna.h
> @@ -51,13 +51,25 @@ struct cna_node {
>  	int			numa_node;
>  	u32			encoded_tail;
>  	u32			pre_scan_result; /* encoded tail or enum val */
> +	u32			intra_count;
>  };
>  
>  enum {
>  	LOCAL_WAITER_FOUND = 2,	/* 0 and 1 are reserved for @locked */
> +	FLUSH_SECONDARY_QUEUE = 3,
>  	MIN_ENCODED_TAIL
>  };
>  
> +/*
> + * Controls the threshold for the number of intra-node lock hand-offs before
> + * the NUMA-aware variant of spinlock is forced to be passed to a thread on
> + * another NUMA node. By default, the chosen value provides reasonable
> + * long-term fairness without sacrificing performance compared to a lock
> + * that does not have any fairness guarantees. The default setting can
> + * be changed with the "numa_spinlock_threshold" boot option.
> + */
> +unsigned int intra_node_handoff_threshold __ro_after_init = 1 << 16;
> +
>  static void __init cna_init_nodes_per_cpu(unsigned int cpu)
>  {
>  	struct mcs_spinlock *base = per_cpu_ptr(&qnodes[0].mcs, cpu);
> @@ -97,6 +109,11 @@ static int __init cna_init_nodes(void)
>  }
>  early_initcall(cna_init_nodes);
>  
> +static __always_inline void cna_init_node(struct mcs_spinlock *node)
> +{
> +	((struct cna_node *)node)->intra_count = 0;
> +}
> +
>  /* this function is called only when the primary queue is empty */
>  static inline bool cna_try_change_tail(struct qspinlock *lock, u32 val,
>  				       struct mcs_spinlock *node)
> @@ -232,7 +249,9 @@ __always_inline u32 cna_pre_scan(struct qspinlock *lock,
>  {
>  	struct cna_node *cn = (struct cna_node *)node;
>  
> -	cn->pre_scan_result = cna_scan_main_queue(node, node);
> +	cn->pre_scan_result =
> +		cn->intra_count == intra_node_handoff_threshold ?
> +			FLUSH_SECONDARY_QUEUE : cna_scan_main_queue(node, node);
>  
>  	return 0;
>  }
> @@ -262,6 +281,9 @@ static inline void cna_pass_lock(struct mcs_spinlock *node,
>  		 * if we acquired the MCS lock when its queue was empty
>  		 */
>  		val = node->locked ? node->locked : 1;
> +		/* inc @intra_count if the secondary queue is not empty */
> +		((struct cna_node *)next_holder)->intra_count =
> +			cn->intra_count + (node->locked > 1);

Playing with lock event counts, I would like you to change the meaning
intra_count parameter that you are tracking. Instead of tracking the
number of times a lock is passed to a waiter of the same node
consecutively, I would like you to track the number of times the head
waiter in the secondary queue has given up its chance to acquire the
lock because a later waiter has jumped the queue and acquire the lock
before it. This value determines the worst case latency that a secondary
queue waiter can experience. So

@@ -332,8 +334,12 @@ static inline void cna_pass_lock(struct
mcs_spinlock *node,
                 */
                val = node->locked ? node->locked : 1;
 
-               /* inc @intra_count if the secondary queue is not empty */
-               next_cn->intra_count = cn->intra_count + (node->locked > 1);
+               /*
+                * inc @intra_count and pass it down if the secondary queue
+                * is not empty
+                */
+               if (node->locked > 1)
+                       next_cn->intra_count = cn->intra_count + 1;
        } else if (node->locked > 1) {    /* if secondary queue is not
empty */
                /* next holder will be the first node in the secondary
queue */

Maybe rename it to jump_count or some other more meaningful name. With
that change, we could probably reduce the default threshold from 64k to
maybe 256 or 512.

I changed the threshold to 256 and run a 96-thread locking stress test
for 10s, the lock event counts:

cna_flush_queue=15687
cna_intra_max=256
cna_mainscan_hit=13
cna_merge_queue=15691
cna_prescan_hit=4344037
cna_prescan_miss=21
cna_splice_new=15701
cna_splice_old=1289
lock_pending=4384
lock_slowpath=47998292
lock_use_node2=16778

Of the prescan hits, only about 0.4% of that resulted in a queue flush
which I thought is reasonable. I didn't see any noticeable degradation
in the performance of the stress test by reducing the threshold from 64k
to 256.

Cheers,
Longman